AGREEMENT AND SENSITIVITY TO BETWEEN-GROUP DIFFERENCES OF U-NET-BASED, AUTOMATED CARTILAGE T2 ANALYSES - DATA FROM THE OA-BIO PROJECT

Abstract

Previous studies reported a high agreement for cartilage segmentations from high resolution 3D gradient echo MRIs using deep learning techniques when compared to manual cartilage segmentations with Dice similarity coefficients (DSC) of up to ∼0.90. Only few studies applied deep learning techniques to multi-echo spin echo (MESE) MRI for cartilage T2 analysis and also reported a somewhat lower agreement (DSC up to 0.75). a) to evaluate the agreement between U-Net-based vs. manual cartilage segmentations of the weight-bearing femorotibial joint using MESE MRIs from Osteoarthritis Initiative (OAI) healthy reference cohort (HRC) participants and b) to evaluate, whether U-Net cartilage segmentations allow to reproduce the differences in cartilage T2 times derived from manual segmentation of a previously conducted matched case-control study of Kellgren & Lawrence grade (KLG) 0 knees with vs. without contralateral (CL) joint space narrowing (JSN). Separate U-Nets were trained for the medial (MFTC) and lateral (LFTC) femorotibial compartment using manual quality-controlled cartilage segmentations from sagittal MESE MRIs of 92 OAI HRC participants (training/validation set n=72/10, 50 epochs, weighted cross-entropy loss function). Training was performed either for all 7 echoes or for the 1st echo only to explore whether the exploitation of the information contained in all echoes provides an advantage over using the 1st echo only. The U-Net segmentation was applied to the 10 remaining HRC knees, the 39 KLG 0 knees with CL JSN (CLJSN) and the matched 39 KLG 0 knees with CL KLG 0 (CLKLG0). The agreement between manual and fully-automated U-Net segmentations was evaluated using the DSC and systematic differences in superficial and deep layer cartilage T2. Cross-sectional and longitudinal cartilage T2 differences were compared between CLJSN and CLKLG0 knees for fully-automated and manual segmentations using Cohen's D as measure of effect size. A high agreement with manual segmentations was observed for both the U-Nets trained on all 7 echoes and the U-Nets trained on the 1st echo only (DSC for both: 0.83 to 0.90). The systematic offset was lower for deep layer T2 (-1.5% to +1.7%) than for superficial layer T2 times (0.2% to 5.1%) and was comparable for the U-Nets trained on all 7 echoes and the U-Nets trained on the 1st echo only. Cross-sectionally, all segmentation methods (manual, all-echoes U-Net, 1st-echo U-Net) showed consistently longer T2 in CLJSN knees than in CLKLG0 knees. In the superficial layer, the Cohen's D for the entire femorotibial joint (FTJ) was greatest for the U-Net trained on all echoes (0.62) followed by the U-Net trained on the 1st echo only (0.57) and the manual segmentations (0.53). In the deep layer, the Cohen's D for the entire FTJ was greatest for the U-Net trained on the 1st echo (0.58) followed by the U-Net trained on all echoes (0.47) and the manual segmentations (0.36). Longitudinally, CLJSN knees tended to show a greater increase in deep layer T2 than CLKLG0 knees. The greatest Cohen's D for deep layer T2 was observed for the entire FTJ with the U-Net trained on all echoes (0.59) followed by manual segmentations (0.50) and the U-Net trained on the 1st echo only (0.47). In the superficial layer, all segmentation methods showed similar, small changes for both CLJSN knees and CLKLG0 knees (Cohen's D≤0.26). The U-Nets applied to MESE MRIs showed high agreement with manual femorotibial cartilage segmentations and only small systematic differences. Training the U-Net on all available echoes had no advantage over training the U-Net on the 1st echo only. When applied to KLG 0 knees with CL JSN vs. CL KLG 0, the U-Net segmentation was able to reproduce the differences previously observed for manual segmentations with a comparable effect size. U-Net-based cartilage segmentations from MESE MRIs may therefore be a promising tool for future studies focusing on cartilage T2 analyses. This work was funded as part of the OA-BIO Eurostars-2 project (E! 114932). AW, SM, FE, WW: Chondrometrics GmbH; FWR: Boston Imaging Core Lab IMI-APPROACH (NCT03883568) participants and investigators CORRESPONDENCE ADDRESS: wolfgang.wirth@pmu.ac.at